I can always use some advice on what to do next.
I have uploaded an image photoset of 1902 photos. When playing with the alignment settings, I managed to align 1885 images in 14 loops and some extra’s.
17 images (not part of the loops but extra individual detail photos) will not register, and when pushing alignment in various ways the 1 big component will scatter in smaller components. I am out of ideas to merge the 17 images in my main component. I tried with extra control points, with changing image setting, but all lead to scattering of my main component. What can I still do to keep my one big component and include some more of the 17 unregistered images?
The 1885 image alignment is made with the following settings:
max features per image 40.000;
image overlap low;
preselector features 20.000;
max reproduction error 2.0
detector sensitivity high;
Group calibration by exif” set to true
The next advice I am looking for is what to do next before the reconstruction? What I have gathered so far on info before the reconstruction phase is, ungroup images when you have your alignment done. I tried setting ungroup images by exif to false and just run alignment again. But this will still scatter the one big component into undesirable smaller ones. So I set group calibration by exif to false again in a new project, and aligned the one big component with ‘merge component only to true’ prior aligning (F6). This made me keep the first big component, but ungrouped (seeing -1 in the image Lens group).
Is this the way to do it right, before starting with the reconstruction phase? Also, to get a better mesh in reconstruction, should I set the max reproduction error to 1.5 or 1 and align again? All I have tried the previous weeks concerns only alignment, I am now up for the reconstruction phase other than preview.
Regards,
Youri
Another opportunity to pose a question that never gets the definitive answer!:
youri_v says “to get [from Alignment] a better mesh in reconstruction …”
Is it right, or not, that once the photos are Aligned, the x + y + z sparse cloud created for Alignment purpose is almost abandoned by RC, and the ‘blobby white’ model is created by the following Reconstruction process, which does it all over again but on a x + y + depth map basis.
So, if that’s correct, it’s not right to expect Alignment’s sparse cloud to help with the quality of the mesh surface?
The aim is to achieve photo alignment by any devious means possible, so then the real work of Reconstruction can happen, which does not rely on or re-use the sparse cloud, but just starts afresh from the Aligned photos, and generates vastly more points?
And also of course, the photos’ images can be fine for the texturing, even if they were majorly inadequate in features for Alignment.
Hi youri,
I’m not sure why you keep asking similar questions in different threads (details not aligning):
https://support.capturingreality.com/hc/en-us/community/posts/360009945371-create-an-all-image-component-helmet-project-
It would be better to report back on suggestions for those similar questions in other threads. Unfortunately, my crystal ball is in the workshop for maintenance at the moment. Next week, when it’s back, I can look in there again to find your answers… 
Whether you need to group your images or ungroup them or in which order is for you to find out. It depends on the object, your desired accuracy and the image set.
Do you mean “Group calibration by exif” in Workflow - Aplication Settings - Import Settings? This, as the location indicates, only influences the result while importing images - it automatically sorts them into calibration groups according to the exif info about the focal length. You can always change that manually or by clicking in 1Ds on “images” (the word, not the “+”), then a secret menu is revealed where you can group and ungroup all of them with one click.
Merge components only is something completely different - it just tells RC not to bother to try and add more left over images to any component, but only to merge already existing components.
You are right, I should stick to one post for follow up. By the way, is your crystal ball out of maintenance yet? I still need some clarification on some things. I have further alligned my helmet images and have done a first reconstruction in normal detail. Took about 50 hours or so, but it was worth the wait. Some details are misaligned though, but I could already see in the point cloud. I am not sure to discart some photos in the misalignment or if I stick with the construct. I have not done the texturing because I want to further align the images for precision. I also want to add some images from previous photo-shoots, because there is too much hight in the top image loops and the middle, creating some less dense spots/ open areas in the reconstruction.
I followed this tutorial:
https://80.lv/articles/full-photogrammetry-guide-for-3d-artists/
and the max feature reprojection error is not that clear to me. My alignment is done with max feature reprojection error of 2.000000. Following this tutorial, you must lower this to 1.000000 for precicion alignment during the ungrouping images. Does this mean I should click on images (the word not the +), click on ungroup, and then decrease the reprojection error before final alignment? Or should I do the final ungrouped images alignment in reprojecetion error 2.000000 and align again after this with ungrouped images with decreased reprojection error?
What I have done so far, when I have my one big component aligned (with grouped images) decreasing the max.reprojection error and align again my component breaks off into many small component. So what is the magic trick to do a precision alignment wih lower reprojection error?
quote “During this step, we can decrease the minimal re-projection error. By default it is 2px. Refine cameras to the at least 1px error, setting it lower will give a more precise alignment in the mean and median errors. Fewer errors, better mesh.”
I tried contacting Vlad Kuzmin, but it seems I can’t send PM to others with this account.
Regards,
Youri
Hi Youri,
yes, how you describe the workflow as it is described is correct in my understanding.
However, 1px max repro error is not a magic trick. It might helpin some cases or not, like in your case. What it does is it tells RC to discard all tie points with an initial repro error of 1 and higher. That means much less to chose from. Raising the value might be worth a try as well.
Anyway, figuring out how to get the best result from your dataset is basically doing the bulk of the work. This is the art of taming RC. So you need to either try it out for yourself with the help of many useful forum-threads that cover many of your questions or you need to provide some more info (see my other posts in your threads) and hope that somebody has some pointers for you…